We investigate use of the anisotropic radial basis functions expansion as a means to represent surface errors on aspheric and freeform surfaces. We show how the optimal choice of the shape parameter and the placement of radial basis function (RBF) nodes can increase accuracy of the surface approximation. We show an example of the adaptive grid refinement. In our approach, complex surfaces are modeled with general arbitrary representation while the anisotropic RBFs expansion models perturbation of the base surface. We show how both the global and the localized surface errors can be modeled across a wide spatial frequency range. With our method, the impact of the structured surface errors on the arbitrary surfaces when applied on the standardized image quality metrics can be assessed for the purpose of optical tolerancing.